Best way to constrain where the attributes for the classes come from?

[torchss]

Assume I have a class with two attributes parameter_1 and parameter_2, that I want outlines to fill out.

Some of my prompts have few-shot examples.

I have noticed that when my context doesn't have meaningful values for parameter_1 and parameter_2, the output from outlines often has values from the examples instead.

It feels like as if the LLM tries its best to provide some values for parameter_1 and parameter_2, even going into its few-shot examples (whereas I need it to pay attention to only the context as mentioned a few times in the prompt)

Currently, I have found a "fix" where I prompt "If you cannot fill this JSON, and only then, answer 'blah' to parameter_1 and 'blah' to parameter_2.", and then if parameter_1 and parameter_2 have 'blah' in them, I know there's no matches in the context.

Another way could have been writing an explicit grammar for the class and then having an "or branch" that says "no match" but before I went down that rabbithole, I wanted to know if there's a better way of handling this

[RobinPicard]

If you're using a model that supports regex as an output type, using Union or Optional as the root of your output type could work.

For instance:

from typing import Optional
from pydantic import BaseModel

class FoundAnswer(BaseModel):
    parameter_1: str
    parameter_2: int

output_type = Optional[FoundAnswer]

# to check whether no answer was found
if response == "None":
    ...

With Union you could similarly define a structure for answer found and another one (a single string or another object, whatever you want) for answer not found.